informs ®

Vol. 35, No. 1, January–February 2005, pp. 49–60 doi 10.1287/inte.1040.0113
issn 0092-2102  eissn 1526-551X  05  3501  0049 © 2005 INFORMS

Bombardier Flexjet Significantly Improves Its
Fractional Aircraft Ownership Operations
Richard Hicks, Richard Madrid, Chris Milligan, Robert Pruneau, Mike Kanaley
Bombardier Aerospace, 3400 Waterview Parkway, Suite 400, Richardson, Texas 75080
{jrhicksjr61@cs.com, rmadrid@flexjet.com, cmilligan@flexjet.com,
robert.pruneau@businessacft.bombardier.com, mkanaley@flexjet.com}

Yvan Dumas, Benoit Lacroix
AD OPT Technologies Inc., 3535, chemin Queen-Mary, Suite 650, Montréal, Quebec, Canada H3V 1H8
{ydumas@ad-opt.com, blacroix@ad-opt.com}

Jacques Desrosiers
HEC Montréal and GERAD, 3000, chemin de la Côte-Sainte-Catherine, Montréal, Quebec, Canada H3T 2A7,
jacques.desrosiers@hec.ca

François Soumis
École Polytechnique de Montréal and GERAD, C.P. 6079, Succursale Centre-ville, Montréal, Quebec, Canada H3C 3A7,
francois.soumis@gerad.ca

The fractional aircraft market is the fastest growing segment of the business aircraft industry. A fractional aircraft
operation is complex—essentially an unscheduled airline in a constantly changing environment. Bombardier
Flexjet implemented a comprehensive three-module optimization system to simultaneously maximize its use
of aircraft, crews, and facilities. AD OPT Technologies designed the modules, using the GENCOL optimizer
developed at GERAD, which employs a column-generation approach to decompose large-scale mixed-integer
nonlinear programming problems. Since inception, the project has generated savings in excess of $54 million
with projected additional savings of $27 million annually, primarily by lowering crew levels (20 percent), aircraft
inventory (40 percent), and supplemental charter aircraft usage (five percent) while increasing aircraft utilization
(10 percent). The quality of customer service has remained consistently high, with significant reduction in
supply.
Key words: transportation: models, network; programming: nonlinear.

T he fractional aircraft market is the fastest grow-
ing sector of the business aircraft industry, with
a 32 percent compounded annual growth rate from
commercial flights. They can also fly anywhere in
the world on as little as 48 hours notice. They pay
predictable monthly management and usage fees,
December 1998 to December 2003. The potential and the operators manage aircraft maintenance, flight
market is estimated to be over 120,000 owners. Frac- crews, hangars, fuel, and insurance on their behalf.
tional ownership programs allow individuals or com- A fractional aircraft ownership operation is complex;
panies to buy shares in a business jet at a fraction of it is essentially an unscheduled airline in a constantly
the full cost of ownership. Owners select the aircraft changing environment with large daily variations in
type appropriate to their needs, determine the num- demand. Operators must match owners’ requests to
ber of hours per year they fly, and purchase shares aircraft and crews in such a way that they operate
of business jets starting at a share size of 1/16th flights safely, satisfy the owners, and maximize prof-
(equal to 50 hours’ flying per year). Firms in the frac- its. They must also schedule crew members and their
tional business guarantee them 24-hour access to a travel, maintain the aircraft, comply with regulations,
fleet with as little as four hours notice. Owners can cope with bad weather, and provide flight-related ser-
fly into 10 times the number of airports available to vices, such as ground transportation and catering.
49

tion (or pairing). ture time windows around the requested times can be bers. and a personalized crew-assignment module for adjusted globally or by individual flight leg. Addi- Bombardier Flexjet’s use of its aircraft. including days with owners purchasing shares can be tax friendly or with heavy demand. Schedulers must con- dollars per live flight hour. mately 80 percent of flights 48 hours or more prior management systems. satisfy owner preferences. Together we began a three-phase scheduling changes in 30 percent of the flights in the system at and optimization enhancement project to maximize least once within 48 hours prior to departure. A rota- rudimentary aircraft routing in reasonable time. Flexjet could benefit from features of the with the proper cost structure and operating charac- phases as we completed them. training. tem were an integrated aircraft and crew optimizer for example. It could obtain only that form a legal workday for a crew member. members may require repositioning by commercial craft.: Bombardier Flexjet Improves Its Fractional Aircraft Ownership Operations 50 Interfaces 35(1). can reduce costs or improve travel service to owners. given that Flexjet must . It could no longer handle that made efficient use of them. an operations to departure and 20 percent with as little lead time research center operated jointly by four Montreal uni. starting and ending at the crew mem- sidering even a small subset of crew working rules ber home base. with about dant (when required). sider many factors not present for scheduled airlines. a its operations using an integer multicommodity flow combination of captain. gain experience with teristics cover days when demand exceeds Flexjet’s them. a leading provider of workforce. contract signings. pp. it optimized of people who operate an aircraft together. and crews. and ground waiting times tions in considering constraints. The crews and aircraft have to crews manually in a suboptimal way. and seek needed revisions as we proceeded with internal supply. empty positioning flights. charter flights). a pref. and flight atten- model with a CPLEX engine. They also pro. We use the following scheduling manually when the fleet size reached terminology throughout the paper: A crew is a group about 25 aircraft. swapped. The overall locations for maintenance vary in maximum capaci- goal was to provide the highest possible number of ties and costs. cover two types of obligations: point-to-point move- duced monthly crew schedules manually using a ments demanded by owners and point-in-time obli- bid-line approach. We also had to integrate the new Managing swaps is important in maximizing the software with the Bombardier Flexjet graphical user utilization of crews. By late 1999. Hard swaps (coming off duty) interface (GUI) to make it widely available to our and soft ones (changing aircraft) are possible. as four hours prior to departure and also ask for versities. In early 2000. The requirements for the new sys. rest periods. and tracked daily separated by days off. 49–60. first officer. Airport limitations prohibit aircraft too on small portions of the scheduling horizon. From 1996 to 2000. There. Con. Multiple reoptimization during daily operations. flight legs. to those who refer prospective owners with cold and warm starts and a repair mode for use to the firm. core air. connections. crews. Hicks et al. costing it on the order of several hundred stand-by duty. Indeed. and to reduce the total opera. that is. and so forth. Multiple locations for signing contracts flawlessly executed flights every day. and GERAD. is a sequence of duties separated by and maintenance restrictions took too long. Swaps among Flexjet crew tional costs (including the costs for crews. Bombardier contracted with AD including the following: Owners request approxi- OPT Technologies. It scheduled The overall integrated system had to allocate aircraft and crews manually for the first year when resources and produce flight and crew assignments its operations were small. its inadequate scheduling gations for which aircraft must be at certain locations system caused Flexjet to maintain extra business jets at certain times for maintenance. and vacation schedule for crew mem. large for an airport’s runways or require extra time erential bidding system for optimizing the monthly buffers between flights because of congestion. © 2005 INFORMS Bombardier entered the fractional aircraft business An Unscheduled Airline in mid-1995 with its Flexjet program. and tions to the fleet of supplemental (charter) aircraft facilities. it realized the model’s limita. The personnel. A monthly block is a sequence of rotations fore operators scheduled. A duty is a combination of 80 aircraft to schedule. crew-swap aspect is complex. Upgrades and downgrades of aircraft the later phases. and supplemental airlines or Flexjet aircraft. Depar- work.

In our project. a long-term planning leg request can alter the schedule of three or four problem. 49–60. TOSCA plans crew (3) Each day. and when and routing. or aircraft (1) Monthly crew assignments based on days off. (6) in case of disrup. and they do not opti. that is. Keskinocak and Tayur (1998) and Martin each day and crews on duty through each overnight et al. For example. period. mization process for using aircraft and crews fol- lows from a sequential division of the problem. con. and forecasted owners’ requests. It tial process and strongly integrated many of its com. considers both hard constraints (based on legality or ponents. ing horizon and plan the movement of all types of mize aircraft itineraries to select good alternative crew aircraft in the fleet. and the number of crews available according to . Matching crew levels to our daily forecasted based models that create single schedules for both targets is important in optimizing aircraft and crews the aircraft and the crew. tude PBS (preferential bidding system) determine ules for crew members. Scheduled airlines (1) build the flight schedule. simultaneous integration of whether to use no crew (unused aircraft). Altitude is a registered trademark of AD ity. that is. For each aircraft. frequently a change in one with the monthly assignments. members’ preferences while considering such events but they have not achieved the strong integration we as crew training. enable Flexjet planners to precisely match crewing Literature on the fractional aircraft business is levels to expected demand in terms of total crews on fairly recent. Furthermore. the scheduled airlines’ opti.to three-day solv- ping only at the end of a duty. we combined the first aircraft. TOSCA determines (2) For all aircraft. The second mod- The scheduled airline’s problem is completely dif. that is. crew rules on top of the aircraft rules. planners to look out over a one. 2001. The system for the Flexjet fractional aircraft physical limitations. (2) assign the fleet. the cen- crew rest periods on the aircraft. one crew flight scheduling. (5) prepare monthly sched. recover aircraft and crews sequentially. Such models TOSCA (the optimizer for scheduling corporate air- are clearly suboptimal because they artificially impose craft) is the heart of the integrated system. company rules. In the second module. major airlines decide the flight commercial airlines when it is used in the repair mode schedule several months in advance while fractional as it recovers aircraft and crews simultaneously. OPT Technologies Inc. annual leaves and monthly schedules (lines). ules and provides personalized schedules. grated optimization modules developed by AD OPT craft. and. struct crew itineraries over several days separately Altitude VBS (vacation bidding system) and Alti- for each type of aircraft. taking into account required maintenance capac. In reengineering the decision process. such as Federal Aviation Admin- business has three modules: istration (FAA) regulations. These long-term planning systems reached for the Flexjet fractional aircraft business. tions. ule also replaces the sequential recovery stages for ferent. that This system architecture is based on the three inte- is. they impose the with TOSCA. cost trade-offs). TOSCA simultaneously assigns swaps. It allows Flexjet one crew duty per aircraft per day and for swap. They allow for only tral short-term optimization module. we started which change constantly. The operators know their flight requests only several hours third module links the results of the previous mod- beforehand. Some they assign crew members duty days and vacations researchers (Cordeau et al. © 2005 INFORMS 51 schedule crew movements around owners’ requests. we modified the usual sequen. assignment of crew members to duty periods considering the crews already assigned working periods on specific aircraft. Mainville Cohn and over four-week periods.: Bombardier Flexjet Improves Its Fractional Aircraft Ownership Operations Interfaces 35(1). (Figure 1). aircraft duty period or two crew duty periods. (2003) present similar multicommodity network. where to swap crew members. often System Architecture resulting in suboptimal decisions.Hicks et al. pp. and crew scheduling. match aircraft types to flights. four stages for commercial airlines. flights to aircraft and optimizes crew duty periods. performance limits) and soft constraints (based on preferences. (3) route the air. trading off coverage require- Barnhart 2003) have tried to integrate aircraft-routing ments (determined by forecasting models) with crew and crew-scheduling modules for commercial airlines. aircraft-type assignment. (4) solve the crew-pairing problem.

It could not prohibit certain crew members from working on cer. TOSCA simultane- ously optimizes aircraft and crew schedules over a one. It considers any constraints that complexity of the scheduling problem. monthly lines determined by Altitude PBS and Alti. and a repair mode in module. COMPASS (crew optimization module for paired A Brief History assignment) assigns individual crew members each The existing Flexjet optimizer with a CPLEX solver day to the crew duty periods that TOSCA has planned (the legacy system) was unable to handle the growing for specific aircraft. The vacation bidding (VBS) and the preferential bidding (PBS) systems determine annual leaves and monthly schedules. two or three minutes or less.to three-day horizon to cover owners’ flight requests.: Bombardier Flexjet Improves Its Fractional Aircraft Ownership Operations 52 Interfaces 35(1). a warm minimize crew-swap costs determined by the TOSCA start in less than 10 minutes. consider many aspects of the aircraft-routing model . Hicks et al. © 2005 INFORMS Demand Forecasts Crew Bids for Days Off. It also matches first officers to captains tude VBS. Once in operation on a day-to-day basis. The TOSCA optimizer can operate in three based on rotation lengths and originating bases to modes: a cold start in less than 30 minutes. pp. It also matches trainers to trainees. 49–60. and COMPASS assigns crew members to specific aircraft and minimizes crew-swap costs. per Day & Aircraft Type Trainings & Vacations Altitude VBS / Altitude PBS Monthly Schedules Crews on Duty Personalized Crew Planning per Day & Aircraft Type (Starting and Ending Days) Customer Requests Crew Assignment to Aircraft (Day t ) Unassigned Crew Bids Available Aircraft Trainings Maintenance Requirements and Capacities Extra Crews TOSCA (Volunteers & Overtime) Commercial Needed Aircraft Crew Swaps Flights for Swaps Charters Routing (When / Where) COMPASS Crew Assignment Crew Swaps to Aircraft (Day t +1 ) (When / Where / Who / How ) Figure 1: The system architecture presents an integrated optimization system that provides routing of aircraft and crews. tain flights. based on long-term forecasts.

it is used to schedule bus required crew swaps to see if by chance they could drivers (Desrochers and Soumis 1989) in such cities as devise a way to swap crews somewhere in the string Tokyo. Singapore. Schedulers had to review systems that are widely used in urban. the Solution Whatever capability the legacy system had to AD OPT Technologies designed the three modules consider. the Flexjet managers A customized branch-and-bound search tree embeds listed the new features they wanted. Modeling all constraints and cost factors for each mod- From the meetings. Because on aver. The optimizer objective . crew constraints did not matter in prac- using the GENCOL optimizer Desrosiers and Soumis tice because Flexjet planners need solutions within developed at GERAD beginning in 1981. ership business and the firm’s optimization problem. UPS. through June 2001). GENCOL: The Mathematics Behind duty rules and the crew-swap requirements. representatives confirmed the technical feasibility. taneously routes and schedules aircraft and crew and daily-crew-assignment optimization software in members. we developed TOSCA. and rail the schedule aircraft by aircraft to find aircraft that transportation. May and June. New York. GENCOL is at the core of several scheduling that need to come off duty. Schedulers a master problem and subproblems alternatively. and they made all sition method that has been adapted to nonlinear the changes after that—both changes to comply with integer programs (Desaulniers et al. For example. © 2005 INFORMS 53 in a reasonable run time. programming algorithms (Desrosiers et al. The solution approach considers duty days or violate overnight crew rules. They awarded the TOSCA following development contracts to AD OPT: simulta. the Flexjet managers determined ule was the main research effort in this project. the schedulers frequently had to change assignments The model used in GENCOL is based on an inte- manually. and Chicago. in particular. preferential-bidding-system tral module is a short-term optimizer that simul- and vacation-bidding-system software in March 2001. GENCOL solves the master problem with a linear pro- Yvan Dumas and Benoit Lacroix from AD OPT Tech. pp. to generate the columns sent to the master prob- sors Jacques Desrosiers and François Soumis from lem. FedEx. departure-time windows. many crew. and Qantas. Bombardier shortest-path problems solved by powerful dynamic Flexjet needed a better system. Columns represent aircraft itineraries and GERAD in March 2000 to explain the fractional own. gram that selects the best mix of columns. that enough new features would be technically fea- sible to justify the expense of the project. During the first phase of the project (April 2000 neous aircraft-and-crew-scheduling optimization soft. used to solve scheduling problems by such airlines as age a third of the fleet requires crew swaps each day. It employs 30 minutes to insure that changes during the run time a column-generation approach to decompose large- do not render the solutions useless. Air Canada. This cen- ware in September 2000. 49–60. In these meetings. it relies on the Dantzig-Wolfe (1960) decompo- and crews for the following day. At the coordinating level. It also optimizes the flight schedule given November 2002. GENCOL nologies later visited the Flexjet owner operations generates only a small subset of the possible billions of center in April and conducted a series of meetings in these aircraft itineraries and crew-member schedules. and the AD OPT this two-level process to obtain integer solutions. crew-member schedules. mathematically decomposes the network-based model comings of their system.Hicks et al.: Bombardier Flexjet Improves Its Fractional Aircraft Ownership Operations Interfaces 35(1). United. 1995) Bombardier representatives first met with Profes. In August. It uses The legacy system did not consider all the basic resource variables and constraints to model the var- crew rules: schedulers had to check all aircraft to ious rules governing aircraft itineraries and crew make sure the optimizer did not schedule overlong members’ schedules. GENCOL owner requests and changes to make up for the short. they authorized the project. that ran the optimizer each morning to schedule flights is. The legacy system was not and places all the nonlinearities within constrained solving the full aircraft-crew problem. The legacy system strung legs together to minimize deadhead (empty) scale mixed-integer nonlinear programming prob- trips without considering which aircraft had crews lems. 1998). air. ger multicommodity network flow problem. It is also of flight legs the optimizer assigned.

maximizing sat- ity of aircraft itineraries and crew members’ sched. that we must not exceed maintenance-location crew member). assignment to different leg types. 49–60. Altitude VBS assigns annual leave to or late departure penalties that discourage our shift. Bombardier Flexjet runs Altitude its to crew members that exceed a heavy workload VBS and Altitude PBS prior to the other modules. score. While satisfying all the craft. during operations. Fleet assignment costs per flight leg. which influence their assignments. grouped together. base to get the new crew. rather than at the end as regular airlines do. willing to work extra days.: Bombardier Flexjet Improves Its Fractional Aircraft Ownership Operations 54 Interfaces 35(1). The early crew planning. days over a one-month period. based on available vacation slots ing departure times from those customers request. © 2005 INFORMS function takes into account costs and penalties. pp. Schedulers must never construct a maximum-score aircraft capabilities for international flights. Re-assignment penal. Bombardier Flexjet’s when and where to swap crews. and crew members initial and of a more senior employee. Hicks et al. each column of the master problem is type of aircraft to assign to each flight request. It flight demand for each day. Altitude PBS determines final conditions. that we must fly flight-leg requests exactly crew legality rules in the subproblems (one for each once. while the right-hand side is the forecasted where to perform maintenance on each aircraft. The local constraints concern the valid. matching customers’ requests that should be of crew swaps while honoring seniority restrictions. annual leaves and training periods. constructing monthly schedules for crew problem constraints into two groups: local and global members consists of assigning days on and off around constraints. 1998). and considers crew members’ preferences. the crew members. and layovers (pairings) for pilots and flight atten- ties when using the optimizer in warm start and dants typically over a one-month period (Gamache repair modes that insure we focus only on near-term et al. the master problem. Altitude PBS assigns work justify using the limited capacity for setup rework. It also determines crew rest including the following: Hourly operating costs for periods and the need for aircraft double-duty crew Flexjet and chartered aircraft. These individual preferences. and from which internal business transformation and systems group . The charter-assignment penalty that In the second phase (April 2001 through April 2002). with specific owners’ requested flights on a daily basis ize crew workloads by applying more restrictive lim. crew rules. During this second phase. For regular airlines. which differ by aircraft type and Altitude VBS and Altitude PBS influence our hierarchies of assignment choices for the various legs. the system ensures coverage of flight capacities and crew availabilities per fleet/base/day. maintains continuity in crew and that we specify maximum aircraft upgrades and rotations. and crew preferences. runway schedule for a particular crew member at the expense limitations. These will be filled Threshold penalties on rest and duty length to equal. Global constraints are imposed on all crew members’ monthly schedules (lines). TOSCA uses column generation to separate the In general. matching the schedules of first officers and flight craft connection times. In summary. The cost for members. It also supports monthly rebids The crew-swap costs that include travel cost per when new vacation slots become available to the crew mile and cost per minute in crew time. TOSCA determines the In GENCOL. Fractional airlines requirements differ changes to the schedule that save enough money to from commercial airlines. time windows on flight depar. isfaction by considering weighted bids that reflect ules and are incorporated in the subproblems. demand with crews. accounts for the losses in quality of service resulting we added optimization for the yearly and monthly from placing owners on non-Flexjet aircraft. the a possible crew schedule that may or may not be aircraft positioning and itineraries. optimizes flight starting times within time windows. Altitude PBS assigns flight legs left uncovered in the solution that must detailed schedules composed of sequences of flights be covered by outside aircraft. and when and selected. that is. aircraft initial and final positions. They the days they have on and off. a set-partitioning model. air. The assignments also favor constraints govern aircraft maintenance times. and it lists those specify that we must select one itinerary per air. downgrades. attendants to captains’ schedules to reduce the costs tures.

between flight segments. In the early stages COMPASS of implementation. Flexjet can thus use the generation of all possible variables (columns) of the system to disseminate good strategies among users master problem. charter hours add up to thousands of dollars. owner. cockpit-qualification. It can exam. flight crews. 49–60. bad outcomes.: Bombardier Flexjet Improves Its Fractional Aircraft Ownership Operations Interfaces 35(1). crew. connection times. pp. matches crew mem. information systems) changes constantly. ule GUI with the GENCOL optimization software an aircraft requires a certain amount of turnaround suite and implemented the COMPASS module. we also want to foster an environ- reserve crew members on standby. FlexSchedule. and days a crew member works). further improved TOSCA and fully integrated it Based on schedulers’ input. and we fre- bers’ rotation ends with aircraft-maintenance sched. on all shifts and move toward the goal of responding . we jet’s other aircraft-scheduling software so the firm encountered resistance to change on the part of the could track all flight crews in real time. As we implemented each new piece of software. we monitored the system closely and asked the schedulers for feedback frequently. COMPASS also considers matching first and features. the feature-rich software in digestible portions. reassignment early flights in the US. to exploit opportunities opened up by new strategies sitioning. and and new processes. For example. Flexjet relies on the sched- members’ preferences to fly with or to avoid flying ulers to see new trends and to work with the devel- with others. the development of the system. It also maximizes crew-member pairings cessful implementation. schedulers who would be using the system and the ine daily crew availability while maintaining full com. COMPASS assigns to prevent aircraft originating outside the US from individual crew members to the crew duty periods being assigned to land at noncustoms airports and TOSCA planned for specific aircraft while considering also restricted these aircraft from assignments from crew-transportation and overtime costs. we designed the system with COMPASS. which it integrated with Flex. optimal. Because the costs for empty positioning and composition. The COMPASS solution method tem to replicate that strategy automatically so that it is essentially a column-generation approach except is applied to the entire schedule and not just when a that the size of the actual problems allows explicit user notices the opportunity. Without their rou- Global constraints on the master problem concern tine feedback. and crew repo. Local constraints on divergence from the optimal operating strategy can the subproblems restrict admissible crew members’ cause major losses. We time before it can be ready for the next flight leg. crew. Flexjet’s operating environ- by home base and rotation length (the consecutive ment (fleet size. We trained schedulers individually. COMPASS obeys any rules that prohibit opers enthusiastically to keep the system current. In the third and final phase (June 2002 through We gathered feedback from them at every stage of December 2003). owner mix. we further integrated the FlexSched- the project. our results could quickly diverge from flights to be covered once. and considers crew operating environment. We seek feedback not just to avoid assignments because of such factors as qualifications. dispatch. the third module. and then we try to adjust the sys- the TOSCA module. We try to notice when users implement officers to captains to reduce crew-swap costs in a strategy manually. matches trainees to trainers. quently review the system to align it with the current ules.Hicks et al. The schedulers were skeptical about new systems FlexSchedule has aircraft. empowering the schedulers. By a crew member from working on a certain flight leg. we enable them to drive We again split the constraints into two categories. owner flying trends. possible extra days. Gathering schedulers’ com- penalties in the warm-start and repair modes. and crew-swap limitations at particu. pliance with FAA regulations for fractional operators. and ments and implementing their suggestions into an assignment of overtime opportunities based on crew adaptive solution methodology was critical to a suc- preferences. we overcame their objections with maintenance displays that helped planners to accept plenty of hands-on training and by implementing and use all the new features of the system. any lar airports or within regions. ment in which we can leverage operations research crew domiciles. © 2005 INFORMS 55 developed an enhanced crew-scheduling-and-tracking Resistance to Change system.

Flexjet’s continues to improve the system based on crew feed. each owner’s departure time. For back. online for the pilots and included recurrent instruc. New Features increasing the size of the problem. by allowing them to bid for five-minute increments and still produce an answer lengths of time on or off without regard to particu. Flexjet designates provided individual instruction in person and over some days as peak days. Flexjet tem could not evaluate delays in a timely manner. about a third of crew members are in tran. It can also swap crews when it is tions and training. By ciently is the make-or-break factor in determining delaying three or four legs slightly. fewer owners need undesirable supplemental longer than one crew’s duty day. also considers the overnight rest requirements. Flexjet also uses the departure window on nonpeak venting assignments that violate overnight rules. Also. so could not consider several elemental concepts. three-hour window manually. This days to allow up to 15-minute delays on any flights. each of those legs the optimizer considers 36 possible delayed departures and 36 possible early departures. schedulers do not have The new automated bidding process replaced the to go through the schedule and make a lot of man- prior manual bid-line system and thus affected the ual changes to accommodate swaps the leg assigner flight crews. It delays. In assigning strings of legs. means the setup is much less elaborate. Since implementation. in less than 30 minutes. We ways depending on the type of day. We made a number of changes after imple. with a five-minute off after a seven-day rotation. on Flexjet’s ables it previously could not. Flexjet discovered it schedulers the opportunity to fix changes by using could delay requested departure times slightly over optimizer warm starts rather than performing only the course of the day without noticeably affecting manual changes. providing During its years of operation. schedulers had to plan any use of the the new integrated system. and Flexjet and so forth up to three hours. the system can assign multiple crews owners as a group: by delaying some flight depar- to one aircraft when it is optimal for it to travel tures. for example. © 2005 INFORMS consistently to all situations regardless of time of day legacy system lacked). 25 percent of the schedule more effectively. This means that. In effect. delays. it considers departure on time. ually and suboptimally. The new optimizer charter aircraft on peak days. Flexjet could avoid whether Flexjet can operate effectively with a reduced using charter aircraft for these. owners may request 190 live legs for a given day. The TOSCA members. We reassured them with communica. it or scheduler. with a 10-minute delay. The system uses the departure-time window in two tion in the crews’ regularly scheduled training. By allowing slight location for crew members from each crew base. Using crew effi. ering the crew-rotation schedule (another input the schedulers would have to make delays manually and . considers the swaps required.: Bombardier Flexjet Improves Its Fractional Aircraft Ownership Operations 56 Interfaces 35(1). optimizer can consider the six-hour time window in mentation. the system can assign approximately 20 to determines which aircraft require swaps by consid. Flexjet now considers vari. it can move owners’ departure times forward or back ously solicited suggestions for new features from crew by up to three hours without penalty. busiest days when optimization was most important. on those days by contract the phone. did not consider. For example. we have continu. for each lar dates and by allowing them to require four days leg. The system establishes inputs have bumped many legs off their assigned company that define the swap time windows for each swap aircraft onto the chartered aircraft. Hicks et al. which would in turn crew-per-aircraft ratio. usually by reduc- sit to or from their assigned aircraft. 49–60. the optimizer can place three or four legs a Each day. ing the time required for positioning empty planes. With for peak days. pre. day on more cost-efficient aircraft. Otherwise. The new optimizer considers all crew duty rules schedulers would effectively build the schedule man- and produces itineraries that honor established con. pp. AD OPT developed the delay. The legacy sys- Because of limitations in the legacy system. We provided complete instructions optimal to do so in the middle of a crew’s rotation. with a 15-minute delay. system with a great deal of crew input. This new feature benefits the straints. By allowing 15-minute A very important new feature controls crew swaps.

For each aircraft that requires maintenance. the growth of the company. Each day members employed to aircraft operated). ized by using a common reference point (for exam- Another new feature that schedulers use every ple. particular location. and the percentage in supple. the key metric for customer service. The sav- Flexjet typically has 10 to 15 maintenance require. For instance. Given the projected rapid growth of the frac- aircraft to maintenance locations. nance personnel can input maintenance events and the annual savings will double. imately 80. We esti- regard to an aircraft’s location or the demand avail. With the legacy system. ings calculated are conservative due to the phased ments to be performed overnight in any of four loca. With savings to grow as the fleet flight hours increase with the multiple-location feature in TOSCA. If the owner base doubles.Hicks et al. By reducing crew levels. which 2003. 400 centage of flights executed flawlessly. mainte. Using ing crew members. Maintenance personnel. It has improved cost measures that influ.000 flight hours. has remained consistently high. who had no training in estimated at a few million dollars. tionality in making our estimate is conservative given the new system replaces those 10 to 15 single-location that the legacy system most likely would break down events with a pool of multiple-location events. levels compared to demand to compete in the frac- ally when the ones entered initially were infeasible.: Bombardier Flexjet Improves Its Fractional Aircraft Ownership Operations Interfaces 35(1). For example. planning for the tional airline business. crew members are getting the number of each aircraft type the location can better schedules than they would have by bidding handle per night. tions. We implemented the preferential bidding dropped. 49–60. Associated with maintenance loca. mated savings at $27 million annually for approx- able to move it around. We estimated that scheduling. for the work AD OPT and Bombardier did can be tion. and we expect annual determined single-location requirements a day. the new sys- tem increased crew utilization (Figure 2). All figures were normal- time to scour the complex schedule for opportunities. senior crew mem- the new system. of schedules were beginning to receive their fifth or ence its scheduling decisions. Using linear propor- the optimizer considers all maintenance bases. The completely long before Bombardier Flexjet doubles its optimizer can use the pool of flights available to get base. we based crew-level savings on the ratio of crew day is the multiple-location maintenance. Crew members per aircraft sold decreased system before most crew members noticed that trend. the total costs legacy system could specify only one particular loca. © 2005 INFORMS 57 would miss opportunities because they had too little was completed in July 2001. Most important. The project has produced many important qualita- tion are maintenance capacity constraints that limit tive benefits. per. core fleet-support aircraft by more than 40 percent. Crew Utilization Flight Hours per Year 600 mental charter aircraft used by more than five per- cent. . the mainte. 300 2000 2001 2002 2003 We based our figures on savings. the For the three phases of the project. Total operations costs sixth choice. pp. our system should prove to be group of aircraft instead of sending each aircraft to a a major enabler of growth for Flexjet. decided which location to use without the return is over $54 million in savings. Thus. Flexjet can alter the operating cost bers who usually received their first or second choice structure. on set lines. a comparison of actual yearly performance until December 2003 with Figure 2: Crew utilization increased by 22 percent between 2000 and conditions before implementation of TOSCA. tional market. implementation. by more than 20 percent. Flexjet needed to reduce its crewing nance personnel scheduled alternate locations manu. while density (live hours per available aircraft 500 per day) has increased by more than 10 percent. That would be 10 to 15 pre. We had a few months’ data under the legacy system after Flex- Financial and Qualitative Impact jet reduced its crew and could see a very strong The integrated system and the new optimization deterioration in the schedules the system was giv- strategies have greatly improved operations.

k∈K ij∈Ak Flexjet has reduced its operations costs. To save space. capacity. its fleet of aircraft k∈K jij∈Ak by more than 40 percent. For each flight leg. The aim of the aircraft-routing problem cost structure. Its scheduling system was the total cost without exceeding the aircraft flying at capacity. For ning of flight j. Flexjet has expanded its product offerings possible to go from the end of flight i to the begin- and increased the flexibility of its owner program. Gk = N k  Ak  over the planning horizon in which Flexjet’s ability to match assets to varying daily N k ⊆ N is the set of flight legs that can be served by demand has improved. The Flexjet Desaulniers et al. a certain . Flexjet determined that it needed to over. and 0 otherwise. Because these savings are proportional to fleet xijk ∈ X k  k ∈ K i j ∈ Ak  (5) size. 49–60. k ∈ K i ∈ N k  (3)  Its total savings in 30 months exceed $50 million. Flexjet was burdened with an uncompetitive of flight j. tijk xijk ≤ mk  k ∈ K (4) Annual savings are about $27 million for 80. example. For leg i ∈ N k  fik is the flying time. © 2005 INFORMS The system helped Flexjet to avoid a major deteriora. This was offered time at ok before the next maintenance. they should grow as the company grows. It was unlikely to be profitable at the is to cover each flight leg once while minimizing asset levels it maintained. Hicks et al. the charter aircraft percent. let cijk management fees. Appendix. increasing their utilization. be the cost incurred (including flight cost at node i on the appropriate type of aircraft k) and tijk be the time Conclusions to go from the beginning of flight i to the beginning In 2000. and we let mk be the remaining flying the industry standard of 50 hours. cost of the aircraft fleet. air- without an increase in the purchase price or monthly craft k flying capacity. Decomposition Method Used in The new system increased efficiency within the GENCOL aircraft-scheduling group by 20 percent and within the crew-planning group by 50 percent. to coordinate maintenance events with flight legs and K a heterogeneous fleet of aircraft maintenance planners. For each arc i j ∈ Ak . pp. and to answer requests made located anywhere. the long-term scheduler could use the simplified version of the decomposition method we time saved to coordinate crew movement with crew use in the TOSCA model.   age by more than five percent. behind GENCOL. min cijk xijk (1) jet has recognized OR/MS as a key strategic tool.000 flight jij∈Ak hours. decreasing   subject to xijk = 1 i ∈ N (2) crew levels by more than 20 percent. (1998) describe the mathematics operations center could focus on other critical areas. that is. so Flexjet will not face capacity limitations for many The objective function in (1) minimizes the total flying years. multicommodity network-flow model: ate new product offerings to continuously adapt to   the rapidly changing fractional airline business. Let N be a set of requested schedulers. and increasing density xijk − xjik = i  jij∈Ak jji∈Ak (live hours per available aircraft per day) by more than 10 percent. Its service level has remained high. construct a graph by owner services. a one-sixteenth share in a Challenger 604 We assume in N k an origin location ok and a fictitious now gives owners 60 hours per year compared with destination d k . a linear- helped Flexjet to achieve operational profitability and programming formulation is given by the following improve its scheduling process. Flexjet can now cre.: Bombardier Flexjet Improves Its Fractional Aircraft Ownership Operations 58 Interfaces 35(1). we describe only a For example. Flex. A Flavor of the Powerful tion in crew morale. We designed the system so that computer speed increases xijk binary k ∈ K i j ∈ Ak  (6) faster than the problem complexity (appendix). For each aircraft. As aircraft k and arc i j ∈ Ak ⊂ N k × N k exists if it is a result. Given that binary flow variables xijk take value 1 if Our three-year project with AD OPT Technologies aircraft k uses arc i j ∈ Ak . haul its asset structure by scheduling more efficiently.

waiting is possible and the has to add various linear and nonlinear functions to starting time of flight j is at least Tik + tijk . This is one of the restrictions that are Additional global constraints can be imposed on this local to the itinerary of aircraft k. the number of con- xij ai ≤ Tik ≤ xijk bik  straints is approximately n. the computer there are as many constraints as there are arcs in all speed increases at most by a factor of eight.Hicks et al. The gen- dk = −1 at the corresponding destination. The size of model (1) through separable by aircraft k and form a block angular struc. the number of variables strained path subproblem for each aircraft while keep- in that model is On3 . the master. In (4). For the subproblems. These additional resource variables of this model becomes harder to solve as the prob- and constraints are just an example of the constraint lem size increases. a linear program.: Bombardier Flexjet Improves Its Fractional Aircraft Ownership Operations Interfaces 35(1). However. As the number of aircraft is a function of the suggests a natural decomposition by defining a con- number of flights. Because flow variables are binary. 49–60. the erated path columns fill in the set-partitioning prob- total flying time of any aircraft must not exceed its lem used to select the right mix of aircraft itineraries. but the num- with (7). For large instances. the above formulation can be solved ties per fleet/base/day are not exceeded and that the directly as an integer linear program. of model (1) through (8) increases by a factor of 32. say n/3. For small instances and maintenance-location capacities and crew availabili- disregarding (5). the number of flights to jij∈Ak jij∈Ak cover. while the number of generated path columns k ∈ K i ∈ N k  (7) that are kept is only about three to five times the number of flights. This the number of aircraft. Tjk is therefore computed as maxakj  Tik + ber of nonzero elements also grows as On2 . If we tijk . mainly depending on constraint set (8). if there is only one large network from which the sys-  k jij∈Ak xij = 1. © 2005 INFORMS 59 aircraft has to cover the leg in (2). Combining take into account all the rules to satisfy. and this is the rea- on the starting time Tik of flight i ∈ N k possibly served son why the direct solution of the linearized version by aircraft k ∈ K. The number of constraints is ing the constraints on covering once each flight leg in also On3 . as long as it does not assume that the number of flights doubles in three exceed bjk . These indicate. These are expressed in used in GENCOL. a piecewise linear function. maximum aircraft upgrade and downgrade assign- straints (3) through (6) and the objective function are ments are specified. a tem chooses the relevant data when solving a specific nonlinear constraint in (8) expresses that if xijk = 1. each of which represents a specific aircraft. this prohibits the number of nonzero elements of the linearized version use of a direct solution approach by linear program. In conclusion. Constraints in (8) can be linearized. For the column-generation process sets that can be defined in (5). Hence. For the      k k  master problem. that ized in (5) by sets X k  k ∈ K. that is. where n is the number of flights. The size of this network is On2 . while the the networks. most of the Time window aki ≤ Tik ≤ bik in (7) is accessible only underlying network information is common. while ok = 1 at the origin of aircraft k. (8) follows: The number of variables is On2  times ture suitable for Dantzig-Wolfe decomposition. that is. we can separate the size of the GENCOL in the form master problem and that of the subproblems. Constraint sets (3) nonlinear shortest-path problems as integer pro- and (6) define a path structure for aircraft k from its grams. pp. but years (at a growth rate of 25 percent). ming and branch and bound. master problem. Observe that the number of nonzero elements in the Consider now imposing time window aki ≤ Tik ≤ bik constraint matrix grows as n5 . This structure is represented by (3) through (8) origin ok to its destination d k : there is conservation together with a modified cost function to account of aircraft flow type when i = 0 for each flight leg for the dual information provided by the master i ∈ N k . a set-partitioning problem. and problem for the pricing of new columns. flying capacity. Others are symbol. con. One then Tjk ≥ Tik + tijk . Therefore. for example. the number of nonzero xijk Tik +tijk −Tjk  ≤ 0 k ∈ K ij ∈ Ak  (8) elements only grows as n2 . if flight i is covered by aircraft k. subproblem. the benefit of the . the number of nonzero elements in the GENCOL performs mathematical decomposition master problem and the subproblems increases only and uses dynamic programming to solve constrained by a factor of four. In GENCOL.

Jacques Daniel Villeneuve from AD OPT for their contributions to Desrosiers. the Desrochers. 49–60. Trans- ations team. Cynthia Barnhart. Michel. Mainville Cohn. Dantzig. 32(3) 246–255. Amsterdam. Irina Ioachim. vice-president of sales at Desrosiers. Cordeau. Éric Gélinas. 51(3) 387–396. Ball. 1960. 35–139. Optimizing Simultaneous assignment of locomotives and cars to passenger on-demand aircraft schedules for fractional aircraft operators. Éric O. berry. Sridhar Tayur. M. Hicks et al. 23(1) 1–13. A column generation entire System Operations team. Nemhauser. Crainic. VBS. approach to the urban transit crew scheduling problem. 49(4) 531–548. Jacques. 2003. the Nether- from AD OPT. Res. Res. Decomposition principle for tial increase in problem size. Daniel Villeneuve. 2001. Elton Gore. for his valuable advice. Improving crew scheduling by incorporating key maintenance routing deci- References sions. Ken Capps. 1995. Solomon. Desaulniers. Cathy Santeiu. We extend our sincere thanks to Nicolas Bélanger and Gamache. portation Sci. Marius M. Monica Fowler. pp. Denton. Ryan Ford. Magnanti. Norwell. 2003. Robert Webb. François Soumis. Guy. François Soumis. François AD OPT. David Jones. Jacques Desrosiers. 8. George B. Transportation Sci. François Lessard from GERAD and Fabrice Lavier Soumis. of the following people at Bombardier Flexjet: Tracy Ash.: Bombardier Flexjet Improves Its Fractional Aircraft Ownership Operations 60 Interfaces 35(1). Pinar. Yvan Dumas. Network Routing.. Transportation Sci. and Yves Siko Handbooks in Operations Research and Management Science. T. Oper. eds. Jason Miller. Elsevier Science B. Gélinas. Fleet Management and Logistics. A uni- fied framework for deterministic time constrained vehicle rout- We gratefully acknowledge the contributions and support ing and crew scheduling problems. L. Martin. 8(1) 101–111. Oper. who developed TOSCA and COMPASS. Finally. G. François Soumis. Daniel Villeneuve. Chris. 32(3) 277–294. MA. Oper. Jacques Desrosiers. . T. 1998. Amy. Arron Miller. Solomon. Res. 1998. 1989. The preferential bidding system GENCOL and to our collaborators Guy Desaulniers and at Air Canada. G. We also acknowledge the important contribu. © 2005 INFORMS decomposition approach is that it allows for substan. linear programs. C. L. eds. Interfaces 33(5) 22–35. Norbert Lingaya. Jean-François. 1998. L.. jet aircraft. who developed Altitude PBS and Altitude lands. François Soumis. our coach for the Edelman presentation. and the entire Flight Oper. Laporte. Kluwer. Michel Gamache at GERAD. Monma. Vol. Gérald Marquis. special thanks go to Keskinocak. tions and efforts of John Allarie. Martin. trains. Time constrained routing and scheduling. Philip Wolfe. 57–93. Scheduling of time-shared Brian T. Pinar Keskinocak.V. from AD OPT. Marius Acknowledgments M.